The diff touches only source, tests, and an append-only entry in docs/DECISIONS.md; AGENTS.md is untouched, so §3.15's reserved-amendment rule is not engaged. Every substantive change tightens rather than loosens a check: `load.ts` now fail-closes a settled research status unless statusNote, statusBy and statusDate are all present; `sourceExists` now refuses to admit a source on a page-title match when it carries an identifier that is unchecked or unmatched (identifier-first admission, consistent with §3.7/§3.8); and TEXT_UNAVAILABLE stops instructing the reader to encode unassessed textual checks as `true` and instead requires `false` plus a stated reason — a direct honesty improvement under §3.8/§3.9. Refused status entries move from stderr-only to a written working file and the run notes, improving inspectability (§3.14/§3.15). The added ResearchCard tooltip attributes a status to the AI edition run that set it and to the panel-judged PR, which is accurate attribution, not a claim of human review (§4). No citations or identifiers are added under content/, and the verification section confirms none to check. The one wrinkle is that the new decisions entry says an AGENTS.md §4 sentence 'is proposed to the founder' while the preceding (base) entry records the founder as already having amended §4; this is stale narration within an append-only log rather than a misattribution of who did what, and does not rise to a violation.

The added decisions-log entry says that the provisional-tier sentence “is proposed to the founder,” but the immediately preceding authoritative entry says that the founder already amended §4 and merged it by hand, and the quoted base constitution already contains that sentence. This creates an internally contradictory history of the amendment’s status, contrary to the requirement that important changes and approvals have accurate, reconstructable provenance. The implementation changes otherwise tighten checks rather than weaken them. Fix the new entry to state that the sentence was proposed and subsequently ratified and merged by the founder, or otherwise give an accurate chronology consistent with the preceding record.

The change responds to prior review notes by tightening verification and provenance requirements. It prevents sources with unresolved identifiers from falling back to page title matching in `sourceExists`, corrects `TEXT_UNAVAILABLE` so unread textual checks report as false rather than true, and enforces fail-closed validation for settled research agenda items (`statusNote`, `statusBy`, and `statusDate`). No constitution files are edited, no citations or records are fabricated, and checks are strengthened rather than weakened.

The diff tightens provisional admission and research-status provenance without editing AGENTS.md: TEXT_UNAVAILABLE now reports unassessed textual checks as false rather than true; sources that carry identifiers may enter only via a resolving match, not a bare page title; settled statuses require statusNote, statusBy, and statusDate fail-closed; refused status rows are written to the run working file and counted in notes; the card surfaces statusBy. DECISIONS.md only records the founder’s already-present §4 provisional tier and these review-note fixes. No citations, confidence material, or check-weakening appear; tests match the stricter gates.

This change tightens rather than weakens the epistemic checks. The verify.ts change (review note #329) closes a real provenance gap: a source carrying an identifier can no longer be admitted on a page title while that identifier is unchecked or unmatched — stricter compliance with §3.8/§4's 'real citations only', with tests covering both branches. The TEXT_UNAVAILABLE rewrite fixes an honesty bug where unassessed textual checks were encoded as `true`: they are now reported `false` with the reason stating they were not assessed, which is exactly the calibrated, honest-labeling behavior §3.13/§4 require. The load.ts change makes a settled research status fail-closed on statusNote, statusBy and statusDate, and refused status entries are now written to the run's working file and counted — disagreement and refusals displayed rather than hidden (§3.15). ResearchCard now names the run that set a status and discloses on hover that it published through the panel-judged PR, improving AI-involvement transparency. Although DECISIONS.md mentions a proposed sentence for AGENTS.md §4, no AGENTS.md edit appears in the diff — the tier it references is already the founder's ratified amendment of 2026-09-17, present in the constitution as quoted. No citations are added, no run records are carried, and the OMITTED FILES list is empty, so nothing unseen could hide a violation.